Deterministic supervisor for agent work

A supervisor daemon for AI coding workflows, replacing an earlier tool that was monolithic and poll-based. Design draws on Marquet's intent-based leadership, Unix process semantics, and a strong preference for deterministic plumbing over LLM orchestration: the supervisor is an event loop, not an agent loop; LLM judgement is summoned as one-shot child processes only at plan-breaking points. State is JSON plus append-only JSONL, no database; checkpoints are git snapshots on hidden refs; verification is external and mechanical (diff plus gate command). His framing: agent boundaries can only offer predicate satisfaction, not reproducibility, so annotate graph edges by contract strength and ask which nodes need to be agentic at all. Builds inside-out, smallest working core first. As of late Sept 2026 he was writing the spec himself and weighing whether to constrain the evaluator to a closed verdict set.

Live: re-check with James before relying on this.

Provenance: James, in conversation with Claude, Sept 2026 (unverified). Disposition: live.

Related: pi-ecosystem, harness-idiom

Part of thread: thread-judgement-in-machines